Robots are trained automatically to recognize and grasp disordered objects
LM (Linemod): Hinterstoisser et al.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes, ACCV 2012, project website, license: CC BY 4.0.
Doris Antensteiner is one of the outstanding researchers at AIT who are requested by external parties for project-related research work. She therefore worked for Siemens for one year as part of the Siemens Residency Research Program.
In her research, she trained robots to determine the position of individual objects and then pick (grasp) them. The challenge was: the objects are disordered and randomly distributed.
Doris Antensteiner (AIT) and her colleagues have developed algorithms that automatically optimize the robot's learning process. Together with Vincent Dietrich and Michael Fiegert, she has published the results in the paper "The Furtherance of Autonomous Engineering via Reinforcement Learning" - and won this year's ICINCO Best Paper Award 2021.
Flexible and dynamic robotics solutions
Dynamic, industrial manufacturing processes lead to an increasing demand for robotic solutions. Until now, the use of industrial robotic systems often required experienced and trained engineers to manually reconfigure the systems for each specific application - in costly and time-consuming processes. Moreover, most systems fail when it comes to reliably locating and recognizing disordered disparate objects and then correctly manipulating them (grasping, handing ...).
In their innovative approach the team around Doris Antensteiner addresses these problems. They implement learning algorithms that train the systems automatically and thus efficiently. In the future, they should help to significantly relieve the burden of classical engineering tasks.
Their solutions can be applied to a wide range of industrial tasks and in different environments, because in the future the robots will "learn" to "adapt" to situations that were previously challenging for them. With the help of the newly developed learning techniques, the systems will be able to autonomously perform picking (gripping) tasks, for example - regardless of the shape and surface structure of the objects. In the future, shadows or reflections will have as little influence on the quality of their work as the partial occlusion or poor illumination of objects.
We congratulate Dorisr on this award and warmly welcome her back to the AIT.
More on
The Furtherance of Autonomous Engineering via Reinforcement Learning
Antensteiner, D.; Dietrich, V. and Fiegert, M. (2021).
In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics - ICINCO, ISBN 978-989-758-522-7; ISSN 2184-2809,
pages 49-59.
DOI: 10.5220/0010544200490059
https://www.scitepress.org/PublicationsDetail.aspx?ID=7W8G07UG9y4=&t=1
Abstract
Engineering efforts are one of the major cost factors in today’s industrial automation systems. We present a configuration system, which grants a reduced obligation of engineering effort. Through self-learning the configuration system can adapt to various tasks by actively learning about its environment. We validate our configuration system using a robotic perception system, specifically a picking application. Perception systems for robotic applications become increasingly essential in industrial environments. Today, such systems often require tedious configuration and design from a well trained technician. These processes have to be carried out for each application and each change in the environment. Our robotic perception system is evaluated on the BOP benchmark and consists of two elements. First, we design building blocks, which are algorithms and datasets available for our configuration algorithm. Second, we implement agents (configuration algorithms) which are designed to intelligently interact with our building blocks. On an examplary industrial robotic picking problem we show, that our autonomous engineering system can reduce engineering efforts.
AIT Center for Vision, Automation & Control
https://www.ait.ac.at/ueber-das-ait/center/center-for-vision-automation-control
AIT researchgroup High-Performance Vision Systems
https://www.ait.ac.at/themen/high-performance-vision-systems
ICINCO Paper Prize Award
http://www.icinco.org
Contact
Dr. DI Mag.ª Doris Antensteiner
ScientistHigh-Performance Vision Systems